Skills or MCP servers: when you need a server

A debate in AI agent development argues that MCP servers are unnecessary for single-user tasks, where simple 'skills'—folders of instructions and scripts—suffice. Skills run in the user's own context with existing credentials, avoiding server overhead. However, MCP servers become essential when multiple users need to share the underlying data.

There’s a take going round that you don’t need MCP servers any more, you just need skills: a folder of instructions and scripts the agent reads and runs. For a lot of jobs that’s right, and it’s usually the simpler answer, so I reach for a skill first. The place it stops being enough is the moment more than one person shares the data behind it. A skill is direction. It’s the prompt, the glue code, the bit of knowledge the model should have and the steps it should follow. And it runs right there in the user’s own context with the user’s own credentials. That’s its strength for a single user: nothing to stand up, nothing to authenticate, it just rides along with whatever access you already have. If the job is “teach the agent how to do this thing” or “wire these two commands together”, a skill is the whole answer and a server would be overhead.